Struct HnswConfig 

pub struct HnswConfig {
    pub m: usize,
    pub m_max0: usize,
    pub ef_construction: usize,
    pub ef_search: usize,
    pub ml: f64,
    pub pq_segments: usize,
    pub pq_centroids: usize,
    pub pq_training_samples: usize,
}

Configuration parameters for an HNSW index.

Parameters

• m - Maximum number of connections per node in each layer. Typical values: 16-64. Higher values give better recall but use more memory.

• m_max0 - Maximum number of connections in layer 0 (the densest layer). Typically set to 2 * m.

• ef_construction - Beam width during index construction. Higher values give better index quality but slower construction. Typical values: 100-500.

• ef_search - Default beam width during search. Higher values give better recall but slower search. Can be overridden per-query. Typical values: 10-500.

• ml - Level multiplier for determining max level. Typically 1 / ln(m). Affects the distribution of nodes across layers.

Product Quantization (PQ) for Compression

When pq_segments is set (non-zero), vectors are compressed using Product Quantization:

• pq_segments - Number of subspaces to divide vectors into. Must divide the vector dimension evenly. Typical values: 8, 16, 32. Higher values = better accuracy but larger codes.

• pq_centroids - Number of centroids per subspace. Default: 256 (8-bit codes).

PQ reduces memory usage by 4-8x with minimal recall loss. During search:

• Full precision vectors are used for the first candidate selection
• Compressed vectors are stored in memory for efficient distance computation
• Original vectors can be retrieved from storage for final reranking

Fields

m: usize

Maximum number of connections per node (M parameter).

m_max0: usize

Maximum connections in layer 0 (typically 2 * M).

ef_construction: usize

Beam width for construction.

ef_search: usize

Default beam width for search.

ml: f64

Level multiplier (1 / ln(M)).

pq_segments: usize

Number of PQ segments (0 = disabled). Must divide vector dimension evenly.

pq_centroids: usize

Number of centroids per PQ segment. Default: 256.

pq_training_samples: usize

Minimum vectors required before PQ training. Default: 1000.

Implementations

impl HnswConfig

pub fn new(m: usize) -> Self

Create a new HNSW configuration with the specified M parameter.

Other parameters are set to sensible defaults:

• m_max0 = 2 * m
• ef_construction = 200
• ef_search = 50
• ml = 1 / ln(m)
• pq_segments = 0 (disabled)

pub const fn with_ef_construction(self, ef: usize) -> Self

Set the beam width for construction.

pub const fn with_ef_search(self, ef: usize) -> Self

Set the default beam width for search.

pub const fn with_m_max0(self, m_max0: usize) -> Self

Set the maximum connections in layer 0.

pub const fn with_pq(self, segments: usize) -> Self

Enable Product Quantization with the specified number of segments.

The vector dimension must be divisible by segments. This reduces memory usage by approximately dimension * 4 / segments bytes per vector.

Arguments

• segments: Number of subspaces. Common values: 8, 16, 32.

Example

// For 128-dim vectors with 8 segments = 16x compression
let config = HnswConfig::new(16).with_pq(8);

pub const fn with_pq_centroids(self, centroids: usize) -> Self

Set the number of centroids per PQ segment.

Default is 256 (8-bit codes). Higher values give better accuracy but require more memory for codebooks.

pub const fn with_pq_training_samples(self, samples: usize) -> Self

Set the minimum number of vectors required before training PQ.

PQ training requires enough samples for k-means clustering. Default is 1000.

pub const fn pq_enabled(&self) -> bool

Check if Product Quantization is enabled.

Trait Implementations

impl Clone for HnswConfig

fn clone(&self) -> HnswConfig

Returns a duplicate of the value.

fn clone_from(&mut self, source: &Self)

Performs copy-assignment from source.

impl Debug for HnswConfig

fn fmt(&self, f: &mut Formatter<'_>) -> Result

Formats the value using the given formatter.

impl Default for HnswConfig

fn default() -> Self

Create a default HNSW configuration.

Uses M=16, which is a good balance between recall and speed.